Automatic Extraction of Music Descriptors from Acoustic Signals Using EDS

By Aymeric Zils and François Pachet.

Proceedings of the 116th Audio Engineering Society Convention, Berlin, Germany, 8–11 May 2004.

Read the paper (PDF)

About the paper

Aymeric Zils and François Pachet describe EDS, the Extractor Discovery System, which constructs audio features automatically for a given description task. Genetic programming combines mathematical and signal-processing operators into candidate functions. Domain knowledge guides the search, and the resulting features are combined into a trained descriptor model.

Abstract summary

The contribution is a method for discovering task-specific signal-processing functions from examples. EDS uses processing patterns, expression simplification and cached computations to make the search practical, then produces an executable extractor. This early work connects automatic feature construction with high-level music description.

Related work

Feature Generation project
Exploring Billions of Audio Features
Analytical Features: A Knowledge-Based Approach to Audio Feature Generation

Back to Publications